Improving within-genus tree species discrimination using the discrete wavelet transform applied to airborne hyperspectral data

Discrete wavelet analysis was assessed for its utility in aiding discrimination of three pine species (Pinus spp.) using airborne hyperspectral data (AVIRIS). Two different sets of Haar wavelet features were compared to each other and to calibrated radiance, as follows: (1) all combinations of detail and final level approximation coefficients and (2) wavelet energy features rather than individual coefficients. We applied stepwise discriminant techniques to reduce data dimensionality, followed by discriminant techniques to determine separability. Leave-one-out cross validation was used to measure the classification accuracy. The most accurate (74.2%) classification used all combinations of detail and approximation coefficients, followed by the original radiance (66.7%) and wavelet energy features (55.1%). These results indicate that application of the discrete wavelet transform can improve species discrimination within the Pinus genus.     

Coniferous tree species mapping using LANDSAT data

The identification and mapping of 12 surface-cover types within Crater Lake National Park, Oregon, including seven classes of coniferous tree species, has been accomplished through the use of LANDSAT digital data. The 12 surface-cover types were mapped with an average accuracy of 88.8%, as compared with detailed ground truth. The classification of the LANDSAT data was accomplished through use of the Interactive Digital Image Manipulation System (IDIMS) available at the EROS Data Center in Sioux Falls, South Dakota. The combined effects of a quantity and quality of ground truth, the use of the controlled clustering classification technique, and the prudent placement of Intensive Study Areas (ISA) on the image-processing CRT screen for training statistic collection provided very subtle spectral reflectance differences between coniferous tree species. Slope angle, slope aspect, and surface-cover type variation, and to a lesser degree, crown size, and crown density were the main environmental factors that accounted for spectral reflectance variation of surface-cover types within the park. Through an appreciation of the influence of environmental factors on the reflectance value of surface-cover types, and an appreciation for the placement of training areas to sample the environmental effects on reflectance, one can reduce misclassification or nonclassification possibilities.     

Object-based fusion for urban tree species classification from hyperspectral,panchromatic and nDSM data

This study aims at identifying the best object-based fusion strategy that takes advantage of the complementarity of several heterogeneous airborne data sources for improving the classification of 15 tree species in an urban area (Toulouse, France). The airborne data sources are: hyperspectral Visible Near-Infrared (160 spectral bands, spatial resolution of 0.4 m) and Short-Wavelength Infrared (256 spectral bands, 1.6 m), panchromatic (14 cm), and a normalized Digital Surface Model (12.5 cm). Object-based feature and decision level fusion strategies are proposed and compared when applied to a reference site where the species are previously identified during ground truth collection. This allows the best fusion strategy to be selected with a view to introducing the method in an automatic process (tree crown delineation and species classification) on a test site, independent of the reference site used for learning. In particular, a decision level fusion is selected: based on the Support Vector Machine algorithm, Visible Near-Infrared and Short-Wavelength Infrared classifications use Minimum Noise Fraction components at the original spatial resolution, whereas panchromatic and normalized Digital Surface Model classifications use, respectively, Haralick’s and structural features computed at the object scale. After the computation of a decision profile for each source at the object level based on the classification algorithms’ membership probabilities, these decision profiles are combined and a decision rule is applied to predict the species. Focusing on the reference site, the Visible Near-Infrared exhibits the best performances with F-score values higher than 60% for 13 species out of 15. The Short-Wavelength Infrared is the most powerful for three species with F-score greater than 60% for seven common species with the Visible Near-Infrared. The panchromatic and normalized Digital Surface Model contribute marginally. The best fusion strategy (decision fusion) does not improve significantly the overall accuracy with 77% (kappa = 74%) against 75% (kappa = 72%) for the Visible Near-Infrared but in general, it improves the results for cases where complementarities have been observed. When applied to the test site and assessed for the two majority species (Tilia tomentosa and Platanus x hispanica), the selected approach gives consistent results with an overall accuracy of 63% against 55% for the Visible Near-Infrared.     

Airborne hyperspectral discrimination of tree species with different ages using discrete wavelet transform

In this article, the capability of discrete wavelet transform (DWT) to discriminate tree species with different ages using airborne hyperspectral remote sensing is investigated. The performance of DWT is compared against commonly used traditional methods, i.e. original reflectance and first and second derivatives. The hyperspectral data are obtained from Thetford forest of the UK, which contains Corsican and Scots pines with different ages and broadleaved tree species. The discrimination is performed by employing three different spectral measurement techniques (SMTs) including Spectral Angle Mapper (SAM), Spectral Information Divergence (SID), and a combination of SAM and SID. Five different mother wavelets with a total of 50 different orders are tested. The wavelet detail coefficient (CD) from each decomposition level and combination of all CDs plus the approximation coefficient from the final decomposition level (C-All) are extracted from each mother wavelet. The results show the superiority of DWT against the reflectance and derivatives for all the three SMTs. In DWT, C-All provided the highest discrimination accuracy compared to other coefficients. An overall accuracy difference of about 20–30% is observed between the finest coefficient and C-All. Amongst the SMTs, SID provided the highest accuracy, while SAM showed the lowest accuracy. Using DWT in combination with SID, an overall accuracy up to around 71.4% is obtained, which is around 13.5%, 14.7%, and 27% higher than the accuracies achieved with reflectance and first and second derivatives, respectively.     

Segmented canonical discriminant analysis of in situ hyperspectral data for identifying 13 urban tree species

A total of 458 in situ hyperspectral data were collected from 13 urban tree species in the City of Tampa, FL, USA using a spectrometer. The 13 species include 11 broadleaf and two conifer species. Three different techniques, segmented canonical discriminant analysis (CDA), segmented principal component analysis (PCA) and segmented stepwise discriminate analysis (SDA), were applied and compared for dimension reduction and feature extraction. With each of the three techniques, 10 features were extracted or selected from four spectral regions, visible (VIS: 1412–1797 nm), near-infrared (NIR: 707–1352 nm), mid-infrared 1 (MIR1: 1412–1797 nm) and mid-infrared 2 (MIR2: 1942–2400 nm), and used to discriminate the 13 urban tree species with a linear discriminate analysis (LDA) method. The cross-validation results, based on training samples that were used in the feature reduction step, and the results calculated from the test samples were used for evaluating the ability of the in situ hyperspectral data and performance of the segmented CDA, PCA and SDA to identify the 13 tree species. The experimental results indicate that a satisfactory discrimination of the 13 tree species was achieved using the segmented CDA technique (average accuracy (AA) = 96%, overall accuracy (OAA) = 96% and kappa = 0.958 from the cross-validation results; AA = 90%, OAA = 90% and kappa = 0.896 from the test samples) compared to the segmented PCA and SDA techniques, respectively (AA = 76% and 86%, OAA = 78% and 87%, and kappa = 0.763 and 0.857 from the cross-validation results; AA = 79% and 88%, OAA = 80% and 89%, and kappa = 0.782 and 0.879 from the test samples). In this study, the segmented CDA transformation is effective for dimension reduction and feature extraction for species discrimination with a relatively limited number of training samples. It outperformed the segmented PCA and SDA methods and produced the highest accuracies. The NIR and MIR1 regions have greater power for identifying the 13 species compared to the VIS and MIR2 spectral regions. The results indicate that CDA or segmented CDA could be applied broadly in mapping forest cover types, species identification and/or other land use/land cover classification practices with hyperspectral remote sensing data.     

The delineation of tree crowns in Australian mixed species forests using hyperspectral Compact Airborne Spectrographic Imager (CASI) data

In mixed-species forests of complex structure, the delineation of tree crowns is problematic because of their varying dimensions and reflectance characteristics, the existence of several layers of canopy (including understorey), and shadowing within and between crowns. To overcome this problem, an algorithm for delineating tree crowns has been developed using eCognition Expert and hyperspectral Compact Airborne Spectrographic Imager (CASI-2) data acquired over a forested landscape near Injune, central east Queensland, Australia. The algorithm has six components: 1) the differentiation of forest, non-forest and understorey; 2) initial segmentation of the forest area and allocation of segments (objects) to larger objects associated with forest spectral types (FSTs); 3) initial identification of object maxima as seeds within these larger objects and their expansion to the edges of crowns or clusters of crowns; 4) subsequent classification-based separation of the resulting objects into crown or cluster classes; 5) further iterative splitting of the cluster classes to delineate more crowns; and 6) identification and subsequent merging of oversplit objects into crowns or clusters. In forests with a high density of individuals (e.g., regrowth), objects associated with tree clusters rather than crowns are delineated and local maxima counted to approximate density. With reference to field data, the delineation process provided accuracies > ∼70% (range 48-88%) for individuals or clusters of trees of the same species with diameter at breast height (DBH) exceeding 10 cm (senescent and dead trees excluded), with lower accuracies associated with dense stands containing several canopy layers, as many trees were obscured from the view of the CASI sensor. Although developed using 1-m spatial resolution CASI data acquired over Australian forests, the algorithm has application elsewhere and is currently being considered for integration into the Definiens product portfolio for use by the wider community.     

基于随机森林的高光谱遥感图像分类

为了充分利用高光谱图像的光谱信息和空间结构信息,提出了一种新的基于随机森林的高光谱遥感图像分类方法,首先,利用主成分分析降低数据的维数,并对主成分进行独立成分分析提取其光谱特征,同时消除像元的空间相关性,再采用形态学分析提取像元的空间结构特征,然后,根据像元的谱域和空域特征分别构造随机森林,并引入空间连续性对像元点的预测结果进行约束修正,最后由投票机制决定最后的分类结果。在AVIRIS和ROSIS高光谱图像上的实验结果表明,所提方法的分类性能要优于传统的高光谱图像分类方法,且分类精度高于基于单一特征的方法。     

Radar modelling of coniferous forest using a tree growth model

The use of a tree growth model to provide statistical information about the microwave scattering components of boreal-type forests (in this case, Scots pine and Norwegian spruce), as an alternative to data obtained through intensive fieldwork, is described. The total backscatter from six test stands at C- and L-band frequency for three polarization combinations (HH, VV and HV) was predicted. Differences between measured C- and L-band data from a polarimetric airborne Synthetic Aperture Radar (EMISAR) and simulated backscatter values compare favourably with previous studies, with like- and cross-polarization differences generally less than 2.5 dB. Modelled backscatter values were consistently less than those observed. A likely explanation for such a discrepancy is the unrealistic manner in which the model incorporates the spatial distribution of tree needles.     

Classification tree analysis using TARGET

Tree models are valuable tools for predictive modeling and data mining. Traditional tree-growing methodologies such as CART are known to suffer from problems including greediness, instability, and bias in split rule selection. Alternative tree methods, including Bayesian CART (Chipman et al., 1998; Denison et al., 1998), random forests (Breiman, 2001a), bootstrap bumping (Tibshirani and Knight, 1999), QUEST (Loh and Shih, 1997), and CRUISE (Kim and Loh, 2001), have been proposed to resolve these issues from various aspects, but each has its own drawbacks.Gray and Fan (2003) described a genetic algorithm approach to constructing decision trees called tree analysis with randomly generated and evolved trees (TARGET) that performs a better search of the tree model space and largely resolves the problems with current tree modeling techniques. Utilizing the Bayesian information criterion (BIC), Fan and Gray (2005) developed a version of TARGET for regression tree analysis. In this article, we consider the construction of classification trees using TARGET. We modify the BIC to handle a categorical response variable, but we also adjust its penalty component to better account for the model complexity of TARGET. We also incorporate the option of splitting rules based on linear combinations of two or three variables in TARGET, which greatly improves the prediction accuracy of TARGET trees. Comparisons of TARGET to existing methods, using simulated and real data sets, indicate that TARGET has advantages over these other approaches.     

Broadleaf species recognition with in situ hyperspectral data

Timely and accurate identification of tree species by spectral methods is crucial for forest and urban ecological management. In this study, a total of 394 reflectance spectra (between 350 and 2500 nm) from foliage branches or canopy of 11 important urban forest broadleaf species were measured in the City of Tampa, Florida, USA with a spectrometer. The 11 species include American elm (Ulmus americana), bluejack oak (Quercus incana), crape myrtle (Lagerstroemia indica), laurel oak (Q. laurifolia), live oak (Q. virginiana), southern magnolia (Magnolia grandiflora), persimmon (Diospyros virginiana), red maple (Acer rubrum), sand live oak (Q. geminata), American sycamore (Platanus occidentalis), and turkey oak (Q. laevis). A total of 46 spectral variables, including normalized spectra, derivative spectra, spectral vegetation indices, spectral position variables, and spectral absorption features were extracted and analysed from the in situ hyperspectral measurements. Two classification algorithms were used to identify the 11 broadleaf species: a nonlinear artificial neural network (ANN) and a linear discriminant analysis (LDA). An analysis of variance (ANOVA) indicates that the 30 selected spectral variables are effective to differentiate the 11 species. The 30 selected spectral variables account for water absorption features at 970, 1200, and 1750 nm and reflect characteristics of pigments and other biochemicals in tree leaves, especially variability of chlorophyll content in leaves. The experimental results indicate that both classification algorithms (ANN and LDA) have produced acceptable accuracies (overall accuracy from 86.3% to 87.8%, kappa from 0.83 to 0.87) and have a similar performance for classifying the 11 broadleaf species with input of the 30 selected spectral variables. The preliminary results of identifying the 11 species with the in situ hyperspectral data imply that with current remote sensing techniques, including high spatial and spectral resolution data, it is still difficult but possible to identify similar species to such 11 broadleaf species with an acceptable accuracy.     

Classification of hyperspectral remote-sensing images using discriminative linear projections

In this study we apply a variant of a recently proposed linear subspace method, the Neighbourhood Component Analysis (NCA), to the task of hyperspectral classification. The NCA algorithm explicitly utilizes the classification performance criterion to obtain the optimal linear projection. NCA assumes nothing about the form of each class and the shape of the separating surfaces. In some cases we would like to weight the penalty function for different types of misclassifications of the algorithm. A modification of the NCA cost function is introduced for this case. Experimental studies are conducted on hyperspectral images acquired by two sensors: the Airborne Visible/Infrared Imaging Spectroradiometer (AVIRIS) and AISA-EAGLE. Experimental results confirm the superiority of the NCA classifier in the context of hyperspectral data classification over methodologies that were previously suggested.     

Daytime fire detection using airborne hyperspectral data

The shortwave infrared region of the electromagnetic spectrum, covering wavelengths from 1400 to 2500 nm, can include significant emitted radiance from fire. There have been relatively few evaluations of the utility of shortwave infrared remote sensing data, and in particular hyperspectral remote sensing data, for fire detection. We used an Airborne Visible InfraRed Imaging Spectrometer (AVIRIS) scene acquired over the 2003 Simi Fire to identify the hyperspectral index that was able to most accurately detect pixels containing fire. All AVIRIS band combinations were used to calculate normalized difference indices, and kappa was used to compare classification ability of these indices for three different fire temperature ranges. The most accurate index was named the Hyperspectral Fire Detection Index (HFDI). The HFDI uses shortwave infrared bands centered at 2061 and 2429 nm. These bands are sensitive to atmospheric attenuation, so the impacts of variable elevation, solar zenith angle, and atmospheric water vapor concentration on HFDI were assessed using radiative transfer modeling. While varying these conditions did affect HFDI values, relative differences between background HFDI and HFDI for 1% fire pixel coverage were maintained. HFDI is most appropriate for detection of flaming combustion, and may miss lower temperature smoldering combustion at low percent pixel coverage due to low emitted radiance in the shortwave infrared. HFDI, two previously proposed hyperspectral fire detection indices, and a broadband shortwave infrared-based fire detection index were applied to AVIRIS scenes acquired over the 2007 Zaca Fire and 2008 Indians Fire. A qualitative comparison of the indices demonstrated that HFDI provides improved detection of fire with less variability in background index values.     

Wildfire temperature and land cover modeling using hyperspectral data

Wildfire temperature retrieval commonly uses measured radiance from a middle infrared channel and a thermal infrared channel to separate fire emitted radiance from the background emitted radiance. Emitted radiance at shorter wavelengths, including the shortwave infrared, is measurable for objects above a temperature of 500 K. The spectral shape and radiance of thermal emission within the shortwave infrared can be used to retrieve fire temperature. Airborne Visible Infrared Imaging Spectrometer (AVIRIS) data were used to estimate fire properties and background properties for the 2003 Simi Fire in Southern California, USA. A spectral library of emitted radiance endmembers corresponding to a temperature range of 500-1500 K was created using the MODTRAN radiative transfer model. A second spectral library of reflected solar radiance endmembers, corresponding to four vegetation types and two non-vegetated surfaces, was created using image spectra selected by minimum endmember average root mean square error (RMSE). The best fit combination of an emitted radiance endmember and a reflected solar radiance endmember was found for each spectrum in the AVIRIS scene. Spectra were subset to reduce the effects of variable column water vapor and smoke contamination over the fire. The best fit models were used to produce maps of fire temperature, fire fractional area, background land cover, land cover fraction, and RMSE. The highest fire temperatures were found along the fire front, and lower fire temperatures were found behind the fire front. Saturation of shortwave infrared channels limited modeling of the highest fire temperatures. Spectral similarity of land cover endmembers and smoke impacted the accuracy of modeled land cover. Sensitivity analysis of modeled fire temperatures revealed that the range of temperatures modeled within 5% of minimum RMSE was smallest between 750 and 950 K. Hyperspectral modeling of wildfire temperature and fuels has potential application for fire monitoring and modeling.     

Application of random forest,generalised linear model and their hybrid methods with geostatistical techniques to count data: Predicting sponge species richness

Spatial distribution of sponge species richness (SSR) and its relationship with environment are important for marine ecosystem management, but they are either unavailable or unknown. Hence we applied random forest (RF), generalised linear model (GLM) and their hybrid methods with geostatistical techniques to SSR data by addressing relevant issues with variable selection and model selection. It was found that: 1) of five variable selection methods, one is suitable for selecting optimal RF predictive models; 2) traditional model selection methods are unsuitable for identifying GLM predictive models and joint application of RF and AIC can select accuracy-improved models; 3) highly correlated predictors may improve RF predictive accuracy; 4) hybrid methods for RF can accurately predict count data; and 5) effects of model averaging are method-dependent. This study depicted the non-linear relationships of SSR and predictors, generated spatial distribution of SSR with high accuracy and revealed the association of high SSR with hard seabed features.     

Classification of melanoma using tree structured wavelet transforms

This paper presents a wavelet transform based tree structure model developed and evaluated for the classification of skin lesion images into melanoma and dysplastic nevus. The tree structure model utilizes a semantic representation of the spatial-frequency information contained in the skin lesion images including textural information. Results show that the presented method is effective in discriminating melanoma from dysplastic nevus. The results are also compared with those obtained using another method of developing tree structures utilizing the maximum channel energy criteria with a fixed energy ratio threshold.     

Using vegetation reflectance variability for species level classification of hyperspectral data

A study was conducted to investigate whether reflectance data from vegetation in a tropical forest canopy could be used for species level discrimination. Reflectance spectra of 11 species were analysed at the scale of the leaf, branch, tree and species. To enhance separation of species-of-interest spectra from the other spectra in the data, the variation in reflectance values for the species-of-interest were used to create a characteristic spectral shape. With a simple algorithm, the resultant shape-space was used as a data filter that correctly discriminated against 94% of the non-species-of-interest trees.     

利用粒子群优化估计高光谱数据协方差矩阵

与传统的多光谱遥感相比,高光谱遥感具有更高的光谱分辨率,能更好地进行地物分类识别。但是,当训练样本数与数据维数相当,或小于后者时,会导致协方差矩阵近似奇异或奇异,使得经典最大似然分类失效,需要对协方差矩阵进行修正。典型的协方差阵估计方法往往只选取总体协方差、类别协方差及其相应变形中的两种形式进行组合,未考虑多种形式共同对协方差阵估计的影响。提出将PSO算法应用到协方差阵估计中,考虑所有形式的共同作用,对组合参数进行优化。最后,通过高光谱数据的分类实验证明了方法的可行性和有效性。     

Noise reduction of hyperspectral data using singular spectral analysis

In this study, a new noise reduction algorithm based on singular spectral analysis (SSA) was developed to reduce the noise in hyperspectral data. With this SSA‐based approach, the reflectance spectrum of a given pixel in a hyperspectral cube is transformed into its state space. The state space is dynamically constructed and characterized by irregular bases, which allows the proposed approach to reduce noises while keeping the absorption features of surface objects. The performance of the developed method was verified on three datasets: two simulated reflectance spectra with several narrow absorption features and a CHRIS (Compact High Resolution Imaging Spectrometer) data cube over agricultural fields. Our results demonstrated the effectiveness of the SSA‐based approach in improving the signal‐to‐noise ratio of hyperspectral data, while keeping the ‘sharp features’ in the reflectance spectra. The results also show that the proposed SSA method outperforms the commonly used MNF (minimum noise fraction) and wavelet‐based noise reduction methods and it improved vegetation cover classification accuracy by 6%.     

Satellite data cloud detection using deep learning supported by hyperspectral data

ABSTRACT

Deep learning methods can play an important role in satellite data cloud detection. The number and quality of training samples directly affect the accuracy of cloud detection based on deep learning. Therefore, selecting a large number of representative and high-quality training samples is a key step in cloud detection based on deep learning. For different satellite data sources, choosing sufficient and high-quality training samples has become an important factor limiting the application of deep learning in cloud detection. This paper presents a fast method for obtaining high-quality learning samples, which can be used for cloud detection of different satellite data with deep learning methods. AVIRIS (Airborne Visible Infrared Imaging Spectrometer) data, which have 224 continuous bands in the spectral range from 400–2500 nm, are used to provide cloud detection samples for different types of satellite data. Through visual interpretation, a sufficient number of cloud and clear sky pixels are selected from the AVIRIS data to construct a hyperspectral data sample library, which is used to simulate different satellite data (such as data from the Moderate Resolution Imaging Spectroradiometer (MODIS) and Landsat Operational Land Imager (OLI) satellites) as training samples. This approach avoids selecting training samples for different satellite sensors. Based on the Keras deep learning framework platform, a backpropagation (BP) neural network is employed for cloud detection from Landsat 8 OLI, National Polar-orbiting Partnership (NPP) Visible Infrared Imaging Radiometer Suite (VIIRS) and Terra MODIS data. The results are compared with cloud coverage results interpreted via artificial vision. The results demonstrate that the algorithm achieves good cloud detection results for the above data, and the overall accuracy is greater than 90%.     

The effect of spectral and spatial degradation of hyperspectral imagery for the Sclerophyll tree species classification

Hyperspectral imaging can be a useful remote-sensing technology for classifying tree species. Prior to the image classification stage, effective mapping endeavours must first identify the optimal spectral and spatial resolutions for discriminating the species of interest. Such a procedure may contribute to improving the classification accuracy, as well as the image acquisition planning. In this work, we address the effect of degrading the original bandwidth and pixel size of a hyperspectral and hyperspatial image for the classification of Sclerophyll forest tree species. A HySpex-VNIR 1600 airborne-based hyperspectral image with submetric spatial resolution was acquired in December 2009 for a native forest located in the foothills of the Andes of central Chile. The main tree species of this forest were then sampled in the field between January and February 2010. The original image spectral and spatial resolutions (160 bands with a width of 3.7 nm and pixel sizes of 0.3 m) were systematically degraded by resampling using a Gaussian model and a nearest neighbour method, respectively (until reaching 39 bands with a width of 14.8 nm and pixel sizes of 2.4 m). As a result, 12 images with different spectral and spatial resolution combinations were created. Subsequently, these images were noise-reduced using the minimum noise fraction procedure and 12 additional images were created. Statistical class separabilities from the spectral divergence measure and an assessment of classification accuracy of two supervised hyperspectral classifiers (spectral angle mapper (SAM) and spectral information divergence (SID)) were applied for each of the 24 images. The best overall and per-class classification accuracies (>80%) were observed when the SAM classifier was applied on the noise-reduced reflectance image at its original spectral and spatial resolutions. This result indicates that pixels somewhat smaller than the tree canopy diameters were the most appropriate to represent the spatial variability of the tree species of interest. On the other hand, it suggests that noise-reduced bands derived from the full image spectral resolution rendered the best discrimination of the spectral properties of the tree species of interest. Meanwhile, the better performance of SAM over SID may result from the ability of the former to classify tree species regardless of the illumination differences in the image. This technical approach can be particularly useful in native forest environments, where the irregular surface of the uppermost canopy is subject to a differentiated illumination.